normalizing flows are an established approach for modelling complex
probability densities through invertible transformations from a base
distribution. However, the accuracy with which the target distribution can be
captured by the normalizing flow is strongly influenced by the topology